Research On Object Detection Algorithm Based On Multi-feature Classification

With the development of technology and the wide application of network communication,people's demand for image processing technology is expanding with the explosive growth of image information,and target detection has become the research priorities of the current machine vision and computer fields.Accurate target detection is a necessary condition for the subsequent intelligent analysis and processing of image recognition,tracking,matching and retrieval.However,in the process of target detection,there are often problems such as complex detection scenes,shadow occlusion,lighting changes,and insufficient use of single feature information.In order to effectively avoid external interference and improve the accuracy of target detection in complex contexts,the target detection algorithm based on multi-feature classification is studied in depth.The main research contents are as follows:(1)The single feature extraction method is studied and analyzed,and we pay more attention on LBP method in texture feature extraction,Canny operator and Sobel operator in edge feature extraction,as well as the related concepts of color histogram and color moment in color feature extraction.At the same time,the sparse theory and the basic theory of support vector machine are analyzed deeply,which is convenient for the research of image preprocessing stage and the improvement of classification algorithm.(2)In the stage of image preprocessing,after studying the noise classification,several common noises and denoising contrast algorithms: Wiener filter,Principal Component Analysis with Local Pixel Grouping(LPG-PCA)and K-SVD,a hybrid image denoising algorithm based on improved SVM and sparse theory is proposed to deal with the noise interference on target detection in complex scenes.In the experimental stage,our method is compared with Wiener filtering denoising,image denoising by using LPG-PCA and K-SVD denoising.The experimental results show that the proposed hybrid scheme has better denoising performance and structural similarity,meanwhile edges and textures are well preserved.(3)After analyzing the advantages of moments,Zernike moments and normalized color histograms are used to extract shape features,texture features and color features.In the feature fusion stage,a multi-kernel learning method is selected to perform feature fusion.In the experimental stage,the recognition rate of single kernel SVM + single descriptor is compared with that of multi-kernel SVM + different combination of feature descriptors,and the time complexity of the algorithm is analyzed.The experimentalresults show that the algorithm has higher recognition rate and faster speed.(4)Inspired by the Particle Swarm Optimization Optimized Invasive Weed Algorithm(IWO-PSO),the advantages of the fireworks algorithm and the characteristics of the PSO algorithm are deeply studied,and the FWA-PSO parameter optimization algorithm is proposed to improve the classification performance of the multi-kernel SVM.Comparing the classification method in this paper with PSO-SVM and FWA-SVM for target detection accuracy,the experimental results show that the parameter optimization algorithm in this paper has the highest detection accuracy.In order to better illustrate the effectiveness of our target detection algorithm,realistic traffic road signs are selected as targets for detection in the final experimental process,and the proposed algorithm is tested in Chinese Traffic Sign Detection Benchmark(CSUST).The experimental results show that the average accuracy rate of the detection results of the proposed algorithm on ban signs,warning signs and indication signs reached 98.77%.